Tether's BotPsy-0.8B Beats 2B Model on Five Agent Tests
TL;DR
- BotPsy-0.8B scores 44.0 on BFCL v4, up from 23.4 for its Qwen3.5-0.8B base, released by Tether AI Research under Apache 2.0.
- The 0.8B model beats Qwen3.5-2B-Think on five of six tool-use and agent benchmarks, trailing only on Claw-Eval (46.5 vs 48.9).
- The card rules it out for safety-critical or high-stakes automated decisions, or tasks involving money, personal information, or health.
A sub-billion-parameter tool-calling model just posted Qwen3.5-2B-level agent scores at a fraction of the size. Tether AI Research released BotPsy-0.8B on Hugging Face under Apache 2.0, a Qwen3.5-0.8B derivative post-trained in four stages: supervised fine-tuning, multi-teacher on-policy distillation, and two rounds of reinforcement learning.
BFCL v4 climbs from 23.4 on the base model to 44.0 on BotPsy. ACEBench, 30.1 to 58.7. τ-bench's retail, telecom and banking subsets, 23.2 to 51.5. The card calls it "Twice the performance of its base model."
The size step is the more unusual claim. "BotPsy has the highest score of all models tested on five of the six benchmarks, beating Qwen3.5-2B-Think, a model 2.5× its size, on BFCL v4 (44.0 vs 37.7), ACEBench (58.7 vs 55.4), τ-bench (51.5 vs 46.5), ToolSandbox (64.6 vs 41.2) and BrowseComp-Plus (4.5 vs 2.8)," the card reports. Qwen3.5-2B-Think wins only on Claw-Eval, 48.9 to 46.5.
The target job is personal-assistant work: "device & system management (settings, files & documents, clocks & alarms), personal productivity & work (calendar & tasks, email, collaboration tools), communication (messaging), information retrieval & research (web search, weather, news), media & lifestyle (video, music, recipes & fitness, photos) and commerce & services (maps & travel, finance, shopping & delivery)."
The limits are blunt. Training and evaluation were primarily in English; the card says BotPsy "performs well in most" of seven other Multi-IF languages "but scores much lower in Hindi and Russian." And: "It is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information, health and other sensitive topics."
Knowledge still scales with parameters. "BotPsy is designed to call tools for facts it does not know, not to serve as a knowledge model," the card notes. BFCL web search sits at 14.0, memory at 16.3, BrowseComp-Plus at 4.5. On τ³-banking, the card observes, no model gets above 5.
A separate Qwen3.5-0.8B agent paper crossed our tracker a day earlier, part of a visible burst of sub-1B agent work.
Originally reported by huggingface.co
Read the original article →Original headline: Tether Ships BotPsy-0.8B, an On-Device Tool-Calling Agent That Beats Models 2.5x Its Size